Apprentice Bot Model Design And Implementation For Psychological Clients’ Therapy

Abstract

In terms of accessibility and reliability, conversational agents employed for advising play an essential role. Conversational bots are not straightforward since different domains may require different algorithms. Extracting information on cases of mental illness, such as consulting clients, the disorder causes treatment and recommendations is challenging in a contemporary mental disorder-related document. Currently, Ethiopia is suffering from a high prevalence of mental health among other developing countries in the world due to poor mental healthcare provision. An apprentice AI text-based chatbot that can consult commonly reported mental disorder cases in Ethiopia for users of the Amharic language has been developed as part of this study. For consulting purposes, this task-oriented bot recognizes features such as general knowledge queries, what to do if a mental disorder happens or is likely to happen, auto-generated consulting advice, and client stories. To construct the bot, the syntactic and semantic structure of data was investigated, as well as a user chat's memory network. To find the best-fitting neural network for the model, four neural networks were experimented with, and existing models were compared. The approach uses word embedding for the semantic extraction of Amharic words and a seq2seq model for suitable response generation. In this thesis, an ensemble architecture of both generative-based and retrieval-based approaches has been employed so that it creates new replies to handle and track user dialogue. The scruffy technique is used to generate word embedding, which is then used to extract semantically comparable terms, to construct unique Amharic word2vec from DSM-5 manual book, articles, and discussion forums obtained from websites. The model has also embedded custom rules for smoothness and error handling during the conversation. The performance and outcomes of the psychological chatbot were demonstrated using perplexity, accuracy measurement, f-1 score metrics, and human-evaluation methods. The model used different model optimization techniques to improve its accuracy. The model's accuracy was 79.62% likelihood of delivering relevant responses at the time of testing, and the model's accuracy was reported as a performance metric. The results demonstrate that for user utterances, the ensemble architecture based on seq2seq modeling with embedded custom rules and Amharic word vector implementation provides pertinent responses.